@echo off
REM Builds WarriorCPU.dll with MSVC. Run this from a plain cmd/PowerShell
REM window - it locates and loads the VS build environment itself, no
REM need to open a "Developer Command Prompt" first.
setlocal

set VSWHERE="%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"
if not exist %VSWHERE% (
    echo Could not find vswhere.exe - is Visual Studio / Build Tools installed?
    exit /b 1
)

for /f "usebackq tokens=*" %%i in (`%VSWHERE% -latest -products * -property installationPath`) do (
    set VSINSTALL=%%i
)
if not defined VSINSTALL (
    echo Could not find any Visual Studio / Build Tools installation.
    exit /b 1
)
if not exist "%VSINSTALL%\VC\Auxiliary\Build\vcvars64.bat" (
    echo Found a VS install at "%VSINSTALL%" but it has no C++ toolset ^(vcvars64.bat missing^).
    echo Install the "Desktop development with C++" workload ^(or the smaller
    echo "Build Tools for Visual Studio"^) and re-run this script.
    exit /b 1
)

call "%VSINSTALL%\VC\Auxiliary\Build\vcvars64.bat"
if errorlevel 1 exit /b 1

cd /d "%~dp0"
REM ---------------------------------------------------------------------------
REM /fp:fast is load-bearing, not cosmetic. Every hot kernel in WarriorCPU.cpp
REM is a floating-point REDUCTION (`sum += in[k] * w[shift+k]` and friends).
REM Under the default /fp:precise the compiler may not reassociate a reduction,
REM so it cannot vectorize one either - the dot product compiles to scalar
REM mulsd/addsd chained through a single accumulator, one double at a time.
REM Measured on the Xeon E5-1650 v2 this trains on, at the real 760-wide
REM geometry: 1.1-1.9 GFLOP/s precise vs 1.7-2.8 GFLOP/s fast, i.e. ~1.5-2x on
REM the three neurons*inputs MAC loops that dominate a training era (forward,
REM hidden gradient, weight-gradient accumulate).
REM Verified safe by DirectML\batch_accum_check.bat: every check passes with
REM byte-identical output to the /fp:precise build, and the deploy-time
REM CPU-vs-MQL5 self-check tolerance (CPU_INFERENCE_MAX_DIFF) is 1.0e-3, some
REM eleven orders of magnitude looser than the drift fast-math can introduce.
REM
REM DO NOT ADD /arch:AVX2. This machine's CPU (Ivy Bridge-EP, 2013) has AVX but
REM NOT AVX2 or FMA, so an AVX2 build turns every kernel into an illegal
REM instruction. That failure is SILENT and looks like success: SehCallFn
REM catches the fault per dispatch, the kernel returns 0, buffers stay
REM untouched, and every shape takes a flat ~5us regardless of size - which
REM benchmarks as a 250x "speedup" until you notice the outputs are all zero.
REM /arch:AVX alone was measured too and bought nothing (these loops are
REM memory-bound, and Ivy Bridge splits 256-bit loads into 2x128 internally).
REM ---------------------------------------------------------------------------
cl.exe /nologo /LD /EHsc /O2 /fp:fast /std:c++17 ^
    WarriorCPU.cpp ^
    /Fe:WarriorCPU.dll

if errorlevel 1 (
    echo Build failed.
    exit /b 1
)

set "TARGET_ROOT=%APPDATA%\MetaQuotes"
if exist "%TARGET_ROOT%\Terminal" (
    for /d %%D in ("%TARGET_ROOT%\Terminal\*") do (
        if exist "%%~fD\MQL5\Libraries" (
            copy /Y "%~dp0WarriorCPU.dll" "%%~fD\MQL5\Libraries\WarriorCPU.dll" >nul 2>&1
        )
    )
)
if exist "%TARGET_ROOT%\Tester" (
    for /d %%D in ("%TARGET_ROOT%\Tester\*") do (
        if exist "%%~fD\MQL5\Libraries" (
            copy /Y "%~dp0WarriorCPU.dll" "%%~fD\MQL5\Libraries\WarriorCPU.dll" >nul 2>&1
        )
    )
)

echo.
echo Built DirectML\WarriorCPU.dll
echo Deployed to any discovered MetaTrader MQL5\Libraries folders.
endlocal